California's No Robo Bosses Act makes employers show workers the data behind AI-led firings

Meta's answer to its employees' lawsuit already uses the language of the new law. The company said the claims "are without merit and not based on facts," and that workforce decisions were made by people, not AI ^[10]. Andon Labs' store experiment followed the same pattern: its AI manager reached a dismissal verdict, and a person made the final decision and delivered the notice ^[13]. A ban on firing based solely on an automated system's judgment ^[2] does not obviously catch either case, because a person signed both times.

So the review step costs little for any employer that already sends terminations past a manager. The law tells the reviewer what to look at, namely work evaluations and performance records ^[3], and those are the records the scoring produced. The Meta plaintiffs argue the system counted time on leave as poor performance ^[11]. All 26 of them had taken protected leave or sought disability accommodations ^[14]. A manager re-reading a rating already lowered by a leave sees the number the model saw. The Andon Labs test shows how far the output depends on the inputs: Luna issued no warning at first because it did not recall the rules, then recommended dismissal once it was given the policy and the record of earlier warnings ^[15].

Luna had an easy case, a worker late on 17 of 23 shifts, or about 74% of them ^[18]. The Meta complaint describes harder ones. A manager was fired 16 days after starting medical leave ^[16]. A scientist got a dismissal notice two days before giving birth on approved leave ^[17].

The second duty is the one that costs money. When AI was a primary tool in the decision, the employer must tell the worker in writing and explain the data used ^[4]. To comply, a company has to be able to reconstruct what fed each score, and the explanation goes to the person most likely to sue. Twenty-six plaintiffs out of about 8,000 cut jobs is roughly 0.3% ^[19]. A notice of that kind would let every dismissed worker check whether leave days entered the score. The law starts July 1 next year ^[5], after the cuts the Meta suit describes ^[20].

Review could become a signature line, leaving the notice as the real expense. Employers could instead argue that AI was only a secondary input, so the notice never triggers; the reporting does not say how the law defines a primary tool. Or regulators could treat a manager's signature on an AI ranking as a decision made solely by the system, and then the review itself gets expensive. I think the first two will happen together: review becomes routine and the dispute moves to the word "primary." That view is wrong if early enforcement treats a nominal sign-off as no review at all.

Newsom signed SB 947 on Sept. 30, according to CNBC as cited by Seoul Economic Daily ^[1], which describes it as the first legal mechanism in the U.S. to require human review of AI personnel decisions ^[22]. It does not stop anyone scoring workers. It assumes AI will flag dismissal candidates and puts a person after it ^[3]. An OECD survey found 90% of U.S. companies use at least one algorithmic management tool ^[8], so the classification question will come up at a large share of employers, or rather at a large share of those with staff in California. The monitoring rules arrive first, in January, with penalties of up to $500 per violation ^[6], about six months ahead of the firing rules ^[21]. Meta had already halted its own monitoring program in June after more than 1,600 employees petitioned against it ^[7].